Automated tapping process for complex parts used for detecting shallow faults

Abstract

In order to carry out the automated tapping of complex parts 10, such as helicopter blades, the surface to be inspected is subject to successive shocks using an impacting head 12 displaced in accordance with a given spacing or pitch. The sound produced by these shocks are collected by a microphone 16 associated with an acquisition circuit 18. At least two successive sound signals, which may or may not be consecutive, are compared in order to establish a fault diagnosis. The comparison and diagnosis are carried out by a neuron network or system, after the number of representative points of each sound signal has been reduced during a preprocessing involving a selection stage followed by a smoothing stage.

What is claimed is: 
     
       1. Process for tapping a complex part, wherein the part is subject to successive shocks based on impacts distributed over a surface to be tested of the part and a diagnosis relative to a presence of shallow faults is deduced from an acoustic response produced by said shocks, in which sound signals produced by each shock are collected and at least two successive sound signals are compared with one another in order to deduce therefrom in automated manner a fault diagnosis when a significant difference is detected between these successive sound signals. 
     
     
       2. Process according to claim 1, wherein calibrated, successive shocks are applied to the part at a regulated frequency and in an automated manner, on the basis of impacts regularly distributed over the surface to be tested. 
     
     
       3. Process according to claim 1, wherein consecutive sound signals produced by the shocks of adjacent impacts are compared. 
     
     
       4. Process according to claim 1, wherein non-consecutive sound signals produced by shocks of impacts separated by at least one intermediate impact are compared. 
     
     
       5. Process according to claim 1, wherein successive sound signals are compared and deduction takes place therefrom in automatic manner of a diagnosis with the aid of a neuron network. 
     
     
       6. Process according to claim 5, wherein q representative values are derived from each of the sound signals during a preprocessing step, q being a positive integer and wherein the q representative values derived from at least two successive sound signals are used as input data of the neuron network. 
     
     
       7. Process according to claim 6, wherein each of the sound signals consists of a sequence of m points, each determined by a time t i  and by a signal amplitude y i  so that the preprocessing consists of a selection stage from among the sequence of m points and p representative points, as well as a smoothing stage, during which the q representative values are deduced from the p representative points, m and p being integers such that m>p>q. 
     
     
       8. Process according to claim 7, wherein the selection stage consists of eliminating the points of the sequence of m points preceding the appearance of a first peak of the sound signal and of only retaining a given number p of points of the sequence of m points as from said first peak. 
     
     
       9. Process according to claim 7, wherein the smoothing stage consists of transforming the times t 1  of each of the p representative points into variables ξ i  such that ξ i  =(2* t i  -Δt)/Δt, in which Δt represents the interval of the variations of the times t i , followed by the determination using the method of least squares of the values of the parameters a for which the expression ##EQU3## is minimal, with n varying from 0 to N, Tc n  (ξ i ) being the degree N Chebyshev's polynomial, said values of parameters a n  being the q sought representative values. 
     
     
       10. Process according to claim 6, wherein the neuron network comprises a four layer neuron network, including an input layer with x * q neurons, x being the number of successive sound signals compared, at least two hidden layers and an output layer having a single neuron. 
     
     
       11. Process according to claim 10, wherein the first hidden layer has an even number of neurons, equal to half the number of neurons of the input layer, the neurons of the first hidden layer being distributed in equal numbers in a first and a second group of neurons, all the neurons of the first group are connected to all the neurons of the x groups of neurons of the input layer, each receiving a first half of the q representative values of one of the compared sound signals and all the neurons of the second group are connected to all the neurons of x groups of neurons of the input layer, each receiving a second half of the q representative values of one of the compared sound signals. 
     
     
       12. Process according to claim 10, wherein the second hidden layer has a number of neurons exceeding 1 and lower than half the number of neurons of the first hidden layer and all the neurons of the second hidden layer are connected to all the neurons of the first hidden layer.